Towards a made-in-Europe ecosystem for multisport training, healthy lifestyle and remote patient monitoring based on cloud-edge continuum of AI-featured body sensors

Towards a made-in-Europe ecosystem for multisport training, healthy lifestyle and remote patient monitoring based on cloud-edge continuum of AI-featured body sensors Luca Borgianni Information Engineering Dep. University of Pisa Pisa, Italy [email protected] Danilo Pietro Pau, FIEEE System Research & Application STMicroelectronics Agrate Brianza, Italy [email protected] Marco Ottella Research & Development Xtremion Technology Lind ob Velden, Austria [email protected] Riccardo Proietti Sport applications Dept. Kubios Oy Kuopio, Finland [email protected] Rudolf Heer Electronic Sensor Systems Silicon Austria Labs Graz, Austria [email protected] Joerg Schotter Electronic Sensor Systems Silicon Austria Labs Graz, Austria [email protected] Juan Montiel-Nelson Inst for appl. Microelectronics Uni. Las Palmas/Gran Canaria Las Palmas/Gran Canaria, Spain j.montiel-[email protected] Mika Tarvainen Department of Technical Physics University of Eastern Finland Kuopio, Finland [email protected] Abstract— In the contemporary landscape of healthcare and fitness, the integration of advanced technological solutions is imperative to address the challenges posed by an ageing society and the increasing demand for personalized, adaptable training programs. This paper delineates the development and potential of a made-in-Europe ecosystem for multisport training, healthy lifestyle, and remote patient monitoring, leveraging the innovative synergy of cloud-edge continuum and AI-enhanced body sensors. The research highlights the critical role of wearable sensors, smart textiles, and TinyML technologies in transforming the accessibility and effectiveness of fitness programs and patient care. Through the deployment of edge computing and SD-WAN, the paper illustrates the seamless integration of real-time data processing, ensuring low latency and high reliability in data transmission. The exploration of heart rate variability (HRV) and breath sensing as tools for personalized training and recovery optimization in endurance sports further exemplifies the application of this integrated approach. Additionally, the challenges and solutions for realtime in-water transmission underscore the potential in aquatic sports monitoring. The paper also presents the advancements in smart textiles, integrating hybrid printed electronics for enhanced user comfort and functionality. Furthermore, the utilization of TinyML and forward learning in human activity recognition demonstrates the capability for real-time, adaptive analytics in sports and healthcare. This comprehensive ecosystem not only promises to revolutionize multisport training and health monitoring but also aims to reduce healthcare costs and improve the quality of life, marking a significant stride towards a holistic, technologically advanced approach in promoting health and fitness across Europe. Keywords— wearable sensors, hearth rate variability, smarttextiles, tinyML, forward-learning, SD-WAN, healthcare, fitness I. INTRODUCTION Ageing society poses challenges for healthcare systems and individuals with age-related health issues. However, a healthy lifestyle with regular physical activity has been proven to be a mitigator of ageing effects, enhancing life quality and longevity. It is also true that accessing personalized, adaptable training programs remains often challenging to the majority of people due to factors like work-life balance and daily commitments. The market today offers a variety of applications, smartwearable devices and sensors for fitness and professional sports activity based on body parameter monitoring, in primis hearth rate and respiration rate. This market sector is huge and steadily growing at 2-digit CAGR, but Europe is not in the leadership. Yet, certain sports, such as swimming, still face technical hurdles. Novel transmission techniques enabled by semiconductors will unleash the potential of real-time monitoring also for water sports. The edge-cloud continuum, seamlessly integrating edge computing and cloud resources, is crucial for wearable sensor success, enabling real-time data processing and reducing latency. Managing multiple connections poses a challenge, but Software-Defined Wide Area Networking (SD-WAN) emerges as a key enabler [1]. SD-WAN optimizes diverse connections, ensuring efficient data routing between wearable devices, local edge servers, and the cloud, enhancing connectivity, adaptability, and reliability in wearable sensor ecosystems. Moreover, integrating these sensors into textiles [2] is crucial for user-friendliness, especially in remote patient monitoring. This paper highlights the emerging technologies in this field, focusing on AI at the edge of nano-sensors and the integration into wearable sensor technology. These innovations are set to foster a holistic ecosystem of hardware, software, and services, supporting a healthy lifestyle with accessible multisport training and remote patient monitoring, thereby reducing hospitalization durations and enhancing post-surgery patient life quality. In this respect the challenge of large-scale semiconductor fabrication (where Europe is a world leader) for healthcare, traditionally limited by low market volume, can be addressed by merging it with the larger fitness/consumer market. Here, by leveraging the European strength in the semiconductor design and fabrication the weakness of the higher levels of the supply chain (Smart Integrated Wearable Systems) will be tackled as well. Finally, this integrated approach will impact the European healthcare sector [3] by reducing costs and burdens. Advanced semiconductors integrated with textiles through advanced wiring processes is poised to revolutionize smart textiles, offering skin compatibility, durability, and resistance to mechanical and chemical stresses. These technologies will enable user-friendly non-invasive, real-time monitoring of complex biochemical parameters, significantly reducing data acquisition errors [4] and improving health and performance monitoring accuracy. II. HEARTH RATE VARIABILITY A. Insights into Heart Rate Variability Heart Rate Variability (HRV) represents the physiological phenomenon of variation in the time interval between heartbeats. It is a non-invasive measure of the autonomic nervous system's (ANS) function and has been extensively studied for its implications in cardiovascular health, stress, and various diseases. HRV is governed by the intricate balance between the sympathetic and parasympathetic branches of the ANS. The variability in heart rate is a manifestation of the body's adaptability to internal and external stimuli. [5] Research activities which was conducted along the last decade, led to the creation of sophisticated software tools and algorithms for HRV analysis. The Kubios HRV software, for instance, is widely used in both clinical and research settings to analyze HRV data as well as in the sport sector [6]. This software delves into the various methodologies including time-domain, frequency-domain, and non-linear analyses and gives very accurate estimations of the recovery status of patients and athletes. This have been instrumental in advancing the understanding of this complex physiological signals providing a foundation for current HRV studies but also paving the way for future advanced algorithms based on AI to unleash the full potential of HRV as a tool for monitoring stress and improving human health and sport performances. B. Applications of HRV in Endurance Sports HRV serves as a critical tool in personalizing training regimens for endurance athletes for several ad-hoc / real-time planning/adaptations. • Training optimization: By monitoring daily HRV scores, coaches and athletes can assess an athlete's readiness for training, ensuring that the intensity and volume of training sessions are aligned with the athlete's current physiological state. The research activities performed at Kubios delves into the nuances of HRV analysis offering a sophisticated approach to interpreting these scores, enabling more precise adjustments to training loads. The readiness monitoring software solution designed for sports coaches is illustrated in Figure 1. • Monitoring recovery: recovery is paramount in endurance sports to prevent overtraining and ensure continuous improvement. HRV provides a quantitative measure of an athlete's recovery status by reflecting the balance between sympathetic and parasympathetic activity. A decrease in HRV, or a shift towards sympathetic dominance, may indicate insufficient recovery. Kubios’ methodologies for HRV analysis can help in distinguishing normal physiological responses from early signs of overtraining syndrome. • Predicting Performance: Emerging research, supported by Kubios’ analytical techniques, suggests that HRV may also serve as a predictor of performance in endurance athletes. Higher HRV values, indicating a well-balanced autonomic nervous system, have been associated with better endurance performance. This relationship underscores the potential of HRV as a non-invasive biomarker for athletic performance. HRV can also help in exercise prescription, since nonlinear correlation properties of HRV have been shown to provide indirect assessment the first and second ventilatory thresholds [7]. • Stress and mental well-being: Endurance sports are not only physically demanding but also mentally taxing. HRV has been used as an indirect marker of psychological stress and mental well-being. Kubios’ HRV analyses of various stress-related conditions can be extended to athletes, offering insights into the psychological aspects of training and competition stress. Managing these stressors is crucial for optimal performance and long-term athlete development. • Biofeedback Training: By learning to control their HRV, athletes can potentially improve their autonomic balance, leading to better stress management, improved recovery, and possibly enhanced performance. This form of training involves real-time HRV monitoring, facilitated by software and analytical tools developed by Kubios and by the integration of them into a edge-cloud continuum system involving sensors, fog nodes and cloud services as described in the following paragraph IV. Figure 1: HRV based athlete’s readiness monitoring solution consisting of: 1) a mobile app, using which athletes can measure daily resting HRV, and 2) a desktop dashboard, using which the coach can monitor the readiness the team and individual athletes. All in all, the applications of HRV in the endurance sports sector are extensive, offering a window into the physiological and psychological state of athletes. Kubios provides a robust framework for leveraging HRV in this context, from optimizing training and recovery to enhancing overall athlete well-being. As the field evolves, integrating HRV monitoring into endurance sports training regimens could become a standard practice, driven by ongoing research and technological advancements. III. REAL-TIME IN-WATER TRANSMISSION CHALLENGES AND SOLUTIONS FOR PHYSICAL ACTIVITY MONITORING The monitoring of underwater activities, presents unique challenges that are distinct from terrestrial or aerial environments. Research activities have been conducted [8] for monitoring of fishes by means of innovative sensor technologies and methodologies. Figure 2 illustrates the developed physical activity and respiration rate tracker. These activities highlighted the critical need for reliable real-time inwater transmission systems, and the main communication challenges that encounters are [9]: signal attenuation, propagation delay, interference and noise. Some innovations and solutions are in literature for overcoming the underwater communication limitations, e.g., acoustic and optical communication, advanced signal processing and hybrid communication systems. Leveraging acoustic or optic signals for data transmission, as they travel better underwater compared to electromagnetic waves, albeit with considerations for speed and bandwidth limitations. Communication with LED (light emitting diodes) has been already proven to be able to transmit in salt water at 10 m range in daylight and with high turbidity with 10 Mb/s rate. [9]. By employing sophisticated signal processing techniques, the detection and decoding of signals in noisy underwater environments is improved. Finally, by combining different communication modalities, such as acoustic, optical, and RF (Radio Frequency) systems, to exploit the advantages of each under specific conditions [10]. However, energy consumption is a major issue, when raw data are transmitted for postprocessing in a cloud. Figure 2: AE-FishBit tracking system for monitoring physical activity and respiration rate. a) Swimming chamber for testing and monitoring of sea bream 50–90 g. b) Activity and breath rate tracker. c) Programming and battery rechargeable unit. These systems are replicable for in-water sport monitoring and this will provide invaluable data for swim trainers and physical therapists to assess the effectiveness and the efficiency of in-water movements and to promptly suggest to the athlete/patients corrective measures. In humans, swimming practice tracking has fewer restrictions than fish tracking, due to the lower weight ratio (device/individual), the larger area available for system deployment and above all that the practice of swimming is done on the surface of the water. Drawing from the state-of the art specific case studies that will be conducted in the next years will assess the practical applications and effectiveness of these solutions in aquatic sport scenarios thus unleashing the potential of remote and real-time accurate bio-mechanical analyses for professional and leisure swimmers as well as for reducing hospitalization time in post-surgery re-habilitation. AI algorithm computation on edge (Edge-AI) by using embedded systems provides the pathway to reduce energy consumption for transmitting activity information to the cloud. In such away, Edge-AI acts as a very-high compression stage before information is transmitted. Edge-AI in combination with advanced RF communication devices, which are integrated in smart textiles, represents the best approach for activity tracker implementation, even in swimming activities. IV. BODY AREA NETWORK AND EDGE-CLOUD CONTINUUM SOLUTIONS A Body Area Network (BAN) represents the paradigm of integrating computing devices and sensors with the human body or connecting them to its surface. BANs incorporate wearable sensors capable of monitoring various physiological parameters, including heart rate, blood pressure, and temperature, with widespread use across various domains, such as healthcare and sports. Leveraging intelligent sensors, BANs can allow the detection and the monitor of various medical conditions, resulting in proactive healthcare management. The proposed architecture aims to integrate AI capabilities both on the Edge and the Cloud, leveraging the novel SDWANtechnology [11] to create a cloud continuum approach. SD-WAN represents a network architecture characterized by a Software-Defined Networking (SDN) approach to manage and optimize wide-area networks (WANs). Unlike traditional WAN solutions that rely on “black-box” hardware and manual configuration, SD-WAN exploits a centralized software controller, enabling dynamic traffic forwarding, application-aware optimization, and centralized management. By leveraging SD-WAN, BANs can establish reliable connections with cloud-based resources, such as healthcare databases and AI algorithms. As shown in Figure 3, we can consider three layers of the proposed integration of SD-WAN in the BAN scenario. BANs consist of wearable sensors transmitting data wirelessly to onboard AI processors for real-time analysis. These sensors represent the Intra-BAN layer. The processed data is then forwarded to a local EDGE Node with AI capabilities that could be considered as a gateway fog device, which acts as an intermediary between the BAN and WAN (Inter-BAN Layer). The gateway fog optimizes data transmission and performs local computational tasks, ensuring efficient utilization of network resources. In the SD-WAN layer of the architecture, our BAN is connected to a heterogeneous network through the utilization of an SD-WAN Customer Premises Equipment (CPE). It serves as a critical component facilitating the connection between the BAN and the SD-WAN-enabled network infrastructure. Specifically, the SD-WAN CPE is deployed at the Edge of the BAN, acting as a gateway to manage the network traffic between the BAN devices and the SD-WAN network. The SD-WAN CPE is responsible for several key functions such as tunnel establishment, traffic forwarding and QoS Management. In this architecture, SD-WAN technology is employed to manage the tunnel selection problem [12] in a heterogeneous network with multiple technologies. When the Edge capabilities are insufficient, or there is an emergency, we need to connect to a Cloud or an emergency service quickly. The integration with an SD-WAN will allow the system to exploit different WAN communication technologies, offering a raliable and cost-effective solution where reliability can be weak when only a single connection technology is considered. Figure 3: Architecture with the integration of BAN and SD-WAN On the other hand, the proposed architecture presents an initial step toward a cloud continuum system for BAN where SD-WAN can forward the traffic to the proper tunnel in real time, aiming to guarantee the desired level of QoS and reliability. V. SMART TEXTILES The integration of hybrid printed electronics into smart textiles represents a significant advancement in the development of wearable smart systems [13]. This integration leverages printing techniques for textile micro-circuit fabrication, enhanced by Nano-plasma processes [14] as well as hybrid electronics printing [15]. Progress was made recently in the field of integration of electronics and sensors directly on textiles by employing different additive manufacturing methods, like inkjet printing or micro-dispensing [16]. Using those technologies, a breath rate sensor was integrated onto a stretchable textile including read out electronics and its performance was evaluated using a tensile stretching machine to mimic human respiration. Figure 4: Integration concept. Taken from [16] A hybrid approach, shown in Figure 4, utilizes digital multilayer inkjet printing technologies to deposit functional materials, needed for integrating electronic components directly onto textiles. Figure 5 a) shows the top view of the seamlessly integrated breath rate sensor device with the piezoresistive strain gauge on a flexible textile. The thermoplastic polyurethane (TPU) encapsulation layer penetrates only through half of the textile fabric, which allows a higher wear comfort. Figure 5 b) shows the two sides of the piezoresistive strain gauge (black lines) on the TPU interposer. The line width and thickness of the dispensed piezoresistive strain gauge were measured with a laser scanning microscope indicated by the blue line in Figure 5 b). The corresponding surface profile is shown in Figure 5 d) with an average line width of 416.6 microns and an average layer thickness of 54.5 microns. Figure 5: a) Top view of the seamlessly integrated breath rate sensor device on textile; b) High-resolution image of the piezoresistive strain gauge on TPU recorded with Laser scanning microscope; c) Fine pitch bonding of breath rate sensor package on the flexible dielectric island using Anisotropic Conductive Film (ACF) technology d) Surface profile of dispensed strain gauge. Taken from [16] These processes are crucial for tailoring thread properties to meet essential criteria such as skin compatibility, wearing comfort, and robustness. Addressing these factors is vital to overcoming existing challenges in wearable technologies, ensuring that the end products are not only functional but also comfortable and appealing to users. To achieve these objectives, further research and innovation (R&I) is necessary, particularly in the early stages of technological development. Engaging end-users from the outset is crucial to align the product development with user needs and preferences, thereby ensuring a higher level of acceptance and opening up broader market opportunities. In the short term, applications such as Remote Assisted Sport Activities and Remote Assisted Living will utilize Electronic components as well as interconnection techniques that are compatible with integration into fabrics [14]. This approach includes the exploration of new flexible, non-fossil materials designed for flexible and structural electronics. These materials are used to create active components, transparent conductors, and barriers, essential for the development of effective wearable devices. Looking towards the mid and long-term future, the focus shifts towards developing large area flexible and stretchable sensors and actuators. This development includes the use of organic and bio-compatible materials, which are pivotal for creating wearable smart systems that combine simultaneous biochemical and biophysical sensing. This comprehensive sensing capability is essential for a wide range of applications, from monitoring health parameters to enhancing athletic performance. The integration of such advanced sensing technologies into textiles paves the way for innovative wearable devices that can seamlessly blend into everyday life, offering unprecedented functionality and user experience. VI. HARNESSING TINYML AND FORWARD LEARNING IN HUMAN ACTIVITY RECOGNITION FOR SPORTS AND HEALTHCARE This paragraph explores how Tiny Machine Learning (Tiny ML) devices, augmented with the capabilities of forward learning on LLMs, are revolutionizing human activity recognition (HAR). TinyML [17] refers to the deployment of machine learning algorithms on low-power, compact devices capable of performing on-device data processing. This technology is especially pertinent in applications requiring real-time analytics without the luxury of high computational power or constant connectivity. In the context of sports and healthcare, TinyML enables wearable devices and embedded systems to identify and monitor various human activities and physiological parameters efficiently [18]. The fusion of TinyML and advanced Large Language Models (LLMs) like GPT-4 offers groundbreaking possibilities in the realms of sport and healthcare. Advancements are proposed frequently to reduce the computational and memory footprint of LLM and Generative AI (GenAI) algorithms. For example [19] proposes deeply quantized 1-bit and 1.58 LLM opening the door for designing a specific hardware capable of taking advantage of low bit-depth mathematical operators. In the same direction [20] proposes LLM-FP4 for quantizing both weights and activations in LLMs down to 4-bit floating-point values, in a post-training manner. These efforts are trying to narrow down the gap between cloud-based implementation and the one at the edge. On the contrary, rather starting from large models and trying to squeeze them to the lowest possible implementation cost without compromising accuracy, very tiny oriented pioneering approaches were performed at STMicroelectronics [21], which has laid the foundation for resource constrained solutions to identifying human activities through compact, efficient devices [22]. To cope with concept drift affecting supervised training, on device forward learning represents a novel approach in machine learning, emphasizing the continuous adaptation and learning of models in dynamic environments [23]. When integrated with LLMs, such as GPT-4, this methodology allows for the incremental absorption of new information, enabling models to evolve and adapt over time without store in a costly manner required data. This feature is particularly beneficial in sports and healthcare, where individual variability and evolving physical conditions necessitate adaptable and personalized models. In sports, the identification and analysis of athlete movements, gestures, and biomechanics are crucial for performance enhancement and injury prevention. TinyML devices, worn by athletes or integrated into sports equipment, can capture a vast array of data points. Forward learning enables these devices to learn from each athlete’s unique patterns, improving the accuracy of activity recognition and providing tailored feedback for performance optimization. 1. Performance Monitoring: Real-time analysis of movements to offer immediate feedback on technique and form. 2. Injury Prevention: Detection of abnormal patterns or fatigue levels that may indicate the risk of injury, allowing for timely intervention. 3. Training Personalization: Adaptation of training regimens based on the individual’s progress and specific physical responses (for instance HRV status as described in paragraph II), promoting optimal performance gains. The healthcare sector benefits immensely from the ability to continuously monitor patients’ physical activities and vital signs outside clinical settings. TinyML devices facilitate nonintrusive monitoring, while forward learning on LLMs ensures the models can adapt to the individual health profiles and conditions of patients, enhancing the precision of care. 1. Remote Patient Monitoring: Continuous tracking of physical activity levels and vital signs for patients with chronic conditions, enabling proactive management of health. 2. Rehabilitation: Monitoring progress in physical rehabilitation programs, with the ability to adjust exercises based on the patient’s recovery trajectory and reducing hospitalization duration. 3. Elderly Care: Detection of falls or unusual inactivity, providing a safety net for the elderly living independently. This technology allows the automatic conversion, optimization, and deployment of pre trained machine learning algorithms into heterogeneous ML capable sensors, microcontrollers (with or without neural hardware accelerators). Therefore, saving precious time that can be spent on innovating the applications. Interoperability with deep learning frameworks is guaranteed thanks to the capability to import standard ML representations. Features like memory allocation, layer folding allows to fit the AI workloads into the scares resources these chips make available. The STM32Cube.AI Developer Cloud [24] service also further add productivity to this process. On top of the ST Edge AI Unified Core technology, sits the Suite. Shown in Figure 7, it was announced at the ST Edge AI Summit on Dec 6, 2023 [25] and it’s a single-entry point to help and support AI developers from idea to deployment on ST products through all intermediate steps. This is to make sure their goals are achieved productively and with very optimal achievements at every stage of their Edge AI workflow. Suite offers adopters guided navigation through a wide choice of free of charge tools. As proof of flexibility and inclusivity of the Suite, Developers enjoy the freedom to “Bring Your Own Data” and “Bring Your Own Model”. The developers will benefit from online documentation, tutorials and be part of a vibrant growing community. ML engineers can use the Suite for any ST device supporting edge AI: STM32 MCUs and MPUs, Smart MEMS Sensors and Stellar Automotive MCUs. Figure 6, ST Unified AI Core Technology Figure 7 ST Edge AI Suite The abovementioned work opens new horizons in human activity recognition. As these technologies evolve, their integration holds the promise of transforming sports performance and healthcare delivery, making personalized, real-time monitoring and intervention a tangible reality. The future of HAR in sports and healthcare looks bright, with TinyML devices and LLMs leading the way towards more intelligent, adaptive, and personalized solutions. VII. CONCLUSIONS This paper has presented an innovative ecosystem of sensors and embedded processors for multisport training, healthy lifestyle, and remote patient monitoring, anchored in the integration of cloud-edge computing and AI-enhanced body sensors within the European context. It has emphasized the pivotal role of wearable sensors and smart textiles, coupled with TinyML technology, in offering personalized and accessible fitness programs and healthcare solutions. The research showcases the utility of heart rate variability (HRV) in customizing training regimens for endurance athletes and explores the challenges of real-time data transmission in aquatic environments. Furthermore, it has discussed the advancements in smart textiles for improved user comfort as well as the application of TinyML and forward learning for efficient human activity recognition. The advancements which are expected in the next years are poised to revolutionize health monitoring and training, promising enhanced quality of life and reduced healthcare costs, thereby contributing significantly to the future of remotely assisted living in the context of fitness and healthcare. ACKNOWLEDGMENT This work is supported by funding from the European Commission through the Horizon Europe Framework Programme and the Chips Joint Undertaking (GA 101140052 and 101130495) as well as from the national funding agencies of Austria, Finland, Germany, Italy, Poland and Spain. REFERENCES [1] L. Borgianni, S. Troia, D. Adami, G. Maier and S. Giordano, "From MPLS to SD-WAN to ensure QoS and QoE in cloud-based applications," 2023 IEEE 9th International Conference on Network Softwarization (NetSoft), Madrid, Spain, 2023, pp. 366-369. [2] T. Agcayazi, K. Chatterjee, A. Bozkurt, and T. K. Ghosh, “Flexible interconnects for electronic textiles,” Advanced Materials Technologies, vol. 3, no. 10, Jan. 2018. [3] M. M. Baig, H. Gholam Hosseini, A. A. Moqeem, F. Mirza, and M. Lindén, “A Systematic review of Wearable Patient Monitoring Systems – Current challenges and Opportunities for clinical adoption,” Journal of Medical Systems, vol. 41, no. 7, Jun. 2017. [4] R. Reddy, R. Pooni, D. P. Zaharieva, B. Senf, J. E. Youssef, E. Dassau, F. J. Doyle, M. A. Clements, M. R. Rickels, S. R. Patton, J. R. Castle, M. C. Riddell, and P. G. Jacobs, “Accuracy of Wrist-Worn activity monitors during common daily physical activities and types of structured exercise: Evaluation study,” Jmir Mhealth and Uhealth, vol. 6, no. 12, p. e10338, Dec. 2018. [5] M. P. Tarvainen, P. O. Ranta-Aho, and P. A. Karjalainen, “An advanced detrending method with application to HRV analysis,” IEEE Transactions on Biomedical Engineering, vol. 49, no. 2, pp. 172–175, Jan. 2002 [6] M. P. Tarvainen, J.-P. Niskanen, J. A. Lipponen, P. O. Ranta-Aho, and P. A. Karjalainen, “Kubios HRV – Heart rate variability analysis software,” Computer Methods and Programs in Biomedicine, vol. 113, no. 1, pp. 210–220, Jan. 2014 [7] B. Rogers and T. Gronwald, “Fractal correlation properties of heart rate variability as a biomarker for intensity distribution and training prescription in endurance exercise: an update”, Front Physiol, 13(Article 879071):1-11, 2022. [8] Martos-Sitcha JA, Sosa J, Ramos-Valido D, Bravo FJ, CarmonaDuarte C, Gomes HL, Calduch-Giner JÀ, Cabruja E, Vega A, Ferrer MÁ, Lozano M, Montiel-Nelson JA, Afonso JM, Pérez-Sánchez J. Ultra-Low Power Sensor Devices for Monitoring Physical Activity and Respiratory Frequency in Farmed Fish. Front Physiol. 2019 May 29;10:667. doi: 10.3389/fphys.2019.00667. PMID: 31191358; PMCID: PMC6548888. [9] A. Caiti, A. Munafò, and R. Petroccia, “Underwater communication,” in Springer eBooks, 2020, pp. 1–10. [10] H. Solar, A. Beriain, R. Berenguer, J. Sosa and J. A. Montiel-Nelson, "Semi-Passive UHF RFID Sensor Tags: A Comprehensive Review," in IEEE Access, vol. 11, pp. 135583-135599, 2023, doi: 10.1109/ACCESS.2023.3336761 [11] L. Borgianni, S. Troia, D. Adami, G. Maier, and S. Giordano, “Assessing the Efficacy of Reinforcement Learning in Enhancing Quality of Service in SD-WANs,” GLOBECOM 2023 - 2023 IEEE Global Communications Conference, Dec. 2023. [12] L. Borgianni, D. Adami, and S. Giordano, “Optimizing Network Performance and Reliability with an Integrated SD-WAN and Satellite 6G Architecture,” 2nd International Conference on 6G Networking (6GNet), Oct. 2023. [13] B. Wang and A. Facchetti, “Mechanically flexible conductors for stretchable and wearable E‐Skin and E‐Textile devices,” Advanced Materials, vol. 31, no. 28, May 2019. [14] R. Alam, G. Hossain, G. Grabher, and Md. M. Hossain, “Applications of low pressure plasma in high-tech textiles,” Linköping Electronic Press Workshop and Conference Collection, Mar. 2013. [15] R. De Fazio, M. Stabile, M. De Vittorio, R. Velázquez, and P. Visconti, “An overview of wearable piezoresistive and inertial sensors for respiration rate monitoring,” Electronics, vol. 10, no. 17, p. 2178, Sep. 2021.. [16] M. Scherf, P. Kulha, G. Kokkinis, R. Heer, P. Stark, Breath Rate Sensor AM Integration Concept for E-textiles, 2023 IEEE SENSORS, 29 October 2023 - 01 November 2023, DOI: 10.1109/SENSORS56945.2023.10324931 [17] https://www.tinyml.org/ visited on 29.02.2024 [18] A. Mazinani, L. Davoli, D. Pau, and G. Ferrari, “Accurate Classification of Sport Activities Under Tiny Deployability Constraints,” IEEE International Conference on Internet of Things and Intelligence Systems (IoTaIS), Nov. 2023. [19] Shuming Ma and Hongyu Wang and Lingxiao Ma and Lei Wang and Wenhui Wang and Shaohan Huang and Li Dong and Ruiping Wang and Jilong Xue and Furu Wei, The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits. 2024, eprint 2402.17764, arXiv, cs.CL. [20] Shih-yang Liu and Zechun Liu and Xijie Huang and Pingcheng Dong and Kwang-Ting Cheng. LLM-FP4: 4-Bit Floating-Point Quantized Transformers, 2023, eprint 2310.16836, arXiv, cs.CL [21] D. Pau and F. M. Aymone, “Forward learning of large language models by consumer devices,” Electronics, vol. 13, no. 2, p. 402, Jan. 2024. [22] D. Pau, P. K. Ambrose, A. Pisani, and F. M. Aymone, “TinyRCE: Forward Learning Under Tiny Constraints,” 2023 IEEE International Conference on Metrology for eXtended Reality, Artificial Intelligence and Neural Engineering (MetroXRAINE), pp. 295-300, Oct. 2023. [23] D. Pau and F. M. Aymone, “Suitability of Forward-Forward and PEPITA Learning to MLCommons-Tiny benchmarks,” 2023 IEEE International Conference on Omni-layer Intelligent Systems (COINS), pp. 1-6, Jul. 2023. [24] https://stm32ai-cs.st.com/home visited on 29.02.2024 [25] https://newsroom.st.com/media-center/press-item.html/c3218.html visited on 29.02.2024